Operator SVD with Neural Networks via Nested Low-Rank Approximation

Authors: Jongha Jon Ryu, Xiangxiang Xu, Hasan Sabri Melihcan Erol, Yuheng Bu, Lizhong Zheng, Gregory W. Wornell

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of the proposed optimization framework for use cases in computational physics and machine learning. Section 4, titled 'Example Applications and Experiments', details empirical evaluations and comparisons with existing methods.
Researcher Affiliation Academia 1Department of EECS, MIT, Cambridge, Massachusetts, USA 2Department of ECE, University of Florida, Gainesville, Florida, USA. Correspondence to: J. Jon Ryu <jongha@mit.edu>.
Pseudocode Yes Section D.1, D.2, and D.4 contain Py Torch code snippets such as 'def get_joint_nesting_masks(...)' and 'class Nested Lo RALoss Function EVD(...)' which describe the method's implementation in a structured, code-like format.
Open Source Code Yes We have opensourced a Py Torch implementation of our method, along with our implementations of Sp IN and Neural EF with a unified I/O interface for a fair comparison.4 The footnote 4 provides the link: 'https://github.com/jongharyu/neural-svd'.
Open Datasets Yes We used the Sketchy Extended dataset (Sangkloy et al., 2016; Liu et al., 2017) to train and evaluate our framework.
Dataset Splits No The paper mentions using a 'standard training setup' and 'test splits' but does not explicitly provide percentages or counts for training, validation, and test dataset splits.
Hardware Specification Yes All experiments were run on a single GPU (NVIDIA Ge Force RTX 3090).
Software Dependencies No The paper mentions using 'Py Torch' for implementation and provides code snippets in Py Torch, but it does not specify the version number for Py Torch or any other software dependencies.
Experiment Setup Yes We trained the networks for 5 105 iterations with batch size 128 and 512. For all methods, we used the RMSProp optimizer (Hinton et al., 2012) with learning rate 10 4 and the cosine learning rate schedule (Loshchilov & Hutter, 2016). For cross-domain retrieval, 'We trained the network for 10 epochs with batch size of 4096. We used the SGD optimizer with learning rate 5 10 3 and momentum 0.9, together with the cosine learning rate schedule'.